Skip to main content

Sourced from docs/quick-start/first-conversation.md in the firmware repo.

Your first conversation

Three ways to talk to the assistant: Telegram, hold-to-talk on the device, and the browser chat.

Telegram (the main channel)​

  1. Create a bot with @BotFather and paste its bot token into Assistant → Connectors → Telegram on the device's web page.
  2. Restart the device (token changes are read at boot).
  3. Message your bot once from your own account. Your message appears on the Telegram card as a pending approval - click Approve. No chat-ID hunting.
  4. Say hello. The assistant replies in text, and, with Voice replies on (the default), can also speak the reply aloud on the device speaker.

The first approved account becomes the owner: only the owner can run the management commands (/update, /loops, /compact, /skill, /help lists them). Everyone else you approve can converse but not manage the device.

Open access

The Open access checkbox lets anyone who finds your bot use it - and your API credits. Leave it off unless you mean to run a public bot.

Hold-to-talk (on the device)​

Press and hold the on-screen mic bar, speak, release. While listening, the Nimbus board's ring breathes red (the all-in-one shows the state on its screen), then a spinner appears while it transcribes; the reply renders on the display (and speaks, when a voice provider that supports the device speaker is configured). Up to 60 seconds per hold.

Everything else is a tap: move between sessions, open the menu, and go back from the touchscreen. The 45-LED ring on the Nimbus board wakes for a glance at live status.

The dashboard tour (two minutes)​

Open the device's web page (scan the sign-in QR on the display if this browser hasn't signed in yet):

  • Dashboard - health, battery, storage, and live sessions. Watch a turn's ring arc mirror here while the assistant works.
  • Chat - a browser conversation with the same assistant, same memory.
  • Assistant - one page with seven subtabs: provider keys and routing (Models), connectors like Telegram (Connectors), the tool surface and Tavily web search (Tools), approved skills (Skills), scheduled routines (Routines), token usage and budgets (Usage), and download trust and guest screening (Safety).
  • Memory & Files - what the assistant remembers (vector memories, conversation history) and every file it has saved; search, preview, share, delete.
  • Routines - scheduled recurring tasks ("morning digest at 8"), each needing your approval before it can run.
  • Settings - battery modes, sound, connectivity, software updates, and the danger zone.

Try asking for something that exercises the machinery: "Remember that my partner's birthday is March 12" (a memory write you can see land in Memory & Files), or "Research the best 2S BMS modules and send me a summary" (a background sub-agent - watch it on the Dashboard).

How it works → Turn anatomy - what the model actually sees